1 //===-- LoopIdiomRecognize.cpp - Loop idiom recognition -------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass implements an idiom recognizer that transforms simple loops into a
11 // non-loop form. In cases that this kicks in, it can be a significant
14 //===----------------------------------------------------------------------===//
18 // Future loop memory idioms to recognize:
19 // memcmp, memmove, strlen, etc.
20 // Future floating point idioms to recognize in -ffast-math mode:
22 // Future integer operation idioms to recognize:
25 // Beware that isel's default lowering for ctpop is highly inefficient for
26 // i64 and larger types when i64 is legal and the value has few bits set. It
27 // would be good to enhance isel to emit a loop for ctpop in this case.
29 // We should enhance the memset/memcpy recognition to handle multiple stores in
30 // the loop. This would handle things like:
31 // void foo(_Complex float *P)
32 // for (i) { __real__(*P) = 0; __imag__(*P) = 0; }
34 // We should enhance this to handle negative strides through memory.
35 // Alternatively (and perhaps better) we could rely on an earlier pass to force
36 // forward iteration through memory, which is generally better for cache
37 // behavior. Negative strides *do* happen for memset/memcpy loops.
39 // This could recognize common matrix multiplies and dot product idioms and
40 // replace them with calls to BLAS (if linked in??).
42 //===----------------------------------------------------------------------===//
44 #define DEBUG_TYPE "loop-idiom"
45 #include "llvm/Transforms/Scalar.h"
46 #include "llvm/ADT/Statistic.h"
47 #include "llvm/Analysis/AliasAnalysis.h"
48 #include "llvm/Analysis/LoopPass.h"
49 #include "llvm/Analysis/ScalarEvolutionExpander.h"
50 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
51 #include "llvm/Analysis/ValueTracking.h"
52 #include "llvm/DataLayout.h"
53 #include "llvm/IRBuilder.h"
54 #include "llvm/IntrinsicInst.h"
55 #include "llvm/Module.h"
56 #include "llvm/Support/Debug.h"
57 #include "llvm/Support/raw_ostream.h"
58 #include "llvm/Target/TargetLibraryInfo.h"
59 #include "llvm/TargetTransformInfo.h"
60 #include "llvm/Transforms/Utils/Local.h"
63 STATISTIC(NumMemSet, "Number of memset's formed from loop stores");
64 STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores");
68 class LoopIdiomRecognize;
70 /// This class defines some utility functions for loop idiom recognization.
73 /// Return true iff the block contains nothing but an uncondition branch
74 /// (aka goto instruction).
75 static bool isAlmostEmpty(BasicBlock *);
77 static BranchInst *getBranch(BasicBlock *BB) {
78 return dyn_cast<BranchInst>(BB->getTerminator());
81 /// Return the condition of the branch terminating the given basic block.
82 static Value *getBrCondtion(BasicBlock *);
84 /// Derive the precondition block (i.e the block that guards the loop
85 /// preheader) from the given preheader.
86 static BasicBlock *getPrecondBb(BasicBlock *PreHead);
89 /// This class is to recoginize idioms of population-count conducted in
90 /// a noncountable loop. Currently it only recognizes this pattern:
92 /// while(x) {cnt++; ...; x &= x - 1; ...}
94 class NclPopcountRecognize {
95 LoopIdiomRecognize &LIR;
97 BasicBlock *PreCondBB;
99 typedef IRBuilder<> IRBuilderTy;
102 explicit NclPopcountRecognize(LoopIdiomRecognize &TheLIR);
106 /// Take a glimpse of the loop to see if we need to go ahead recoginizing
108 bool preliminaryScreen();
110 /// Check if the given conditional branch is based on the comparison
111 /// beween a variable and zero, and if the variable is non-zero, the
112 /// control yeilds to the loop entry. If the branch matches the behavior,
113 /// the variable involved in the comparion is returned. This function will
114 /// be called to see if the precondition and postcondition of the loop
115 /// are in desirable form.
116 Value *matchCondition (BranchInst *Br, BasicBlock *NonZeroTarget) const;
118 /// Return true iff the idiom is detected in the loop. and 1) \p CntInst
119 /// is set to the instruction counting the pupulation bit. 2) \p CntPhi
120 /// is set to the corresponding phi node. 3) \p Var is set to the value
121 /// whose population bits are being counted.
123 (Instruction *&CntInst, PHINode *&CntPhi, Value *&Var) const;
125 /// Insert ctpop intrinsic function and some obviously dead instructions.
126 void transform (Instruction *CntInst, PHINode *CntPhi, Value *Var);
128 /// Create llvm.ctpop.* intrinsic function.
129 CallInst *createPopcntIntrinsic(IRBuilderTy &IRB, Value *Val, DebugLoc DL);
132 class LoopIdiomRecognize : public LoopPass {
134 const DataLayout *TD;
137 TargetLibraryInfo *TLI;
138 const ScalarTargetTransformInfo *STTI;
141 explicit LoopIdiomRecognize() : LoopPass(ID) {
142 initializeLoopIdiomRecognizePass(*PassRegistry::getPassRegistry());
143 TD = 0; DT = 0; SE = 0; TLI = 0; STTI = 0;
146 bool runOnLoop(Loop *L, LPPassManager &LPM);
147 bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
148 SmallVectorImpl<BasicBlock*> &ExitBlocks);
150 bool processLoopStore(StoreInst *SI, const SCEV *BECount);
151 bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount);
153 bool processLoopStridedStore(Value *DestPtr, unsigned StoreSize,
154 unsigned StoreAlignment,
155 Value *SplatValue, Instruction *TheStore,
156 const SCEVAddRecExpr *Ev,
157 const SCEV *BECount);
158 bool processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
159 const SCEVAddRecExpr *StoreEv,
160 const SCEVAddRecExpr *LoadEv,
161 const SCEV *BECount);
163 /// This transformation requires natural loop information & requires that
164 /// loop preheaders be inserted into the CFG.
166 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
167 AU.addRequired<LoopInfo>();
168 AU.addPreserved<LoopInfo>();
169 AU.addRequiredID(LoopSimplifyID);
170 AU.addPreservedID(LoopSimplifyID);
171 AU.addRequiredID(LCSSAID);
172 AU.addPreservedID(LCSSAID);
173 AU.addRequired<AliasAnalysis>();
174 AU.addPreserved<AliasAnalysis>();
175 AU.addRequired<ScalarEvolution>();
176 AU.addPreserved<ScalarEvolution>();
177 AU.addPreserved<DominatorTree>();
178 AU.addRequired<DominatorTree>();
179 AU.addRequired<TargetLibraryInfo>();
182 const DataLayout *getDataLayout() {
183 return TD ? TD : TD=getAnalysisIfAvailable<DataLayout>();
186 DominatorTree *getDominatorTree() {
187 return DT ? DT : (DT=&getAnalysis<DominatorTree>());
190 ScalarEvolution *getScalarEvolution() {
191 return SE ? SE : (SE = &getAnalysis<ScalarEvolution>());
194 TargetLibraryInfo *getTargetLibraryInfo() {
195 return TLI ? TLI : (TLI = &getAnalysis<TargetLibraryInfo>());
198 const ScalarTargetTransformInfo *getScalarTargetTransformInfo() {
200 TargetTransformInfo *TTI = getAnalysisIfAvailable<TargetTransformInfo>();
201 if (TTI) STTI = TTI->getScalarTargetTransformInfo();
206 Loop *getLoop() const { return CurLoop; }
209 bool runOnNoncountableLoop();
210 bool runOnCountableLoop();
214 char LoopIdiomRecognize::ID = 0;
215 INITIALIZE_PASS_BEGIN(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
217 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
218 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
219 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
220 INITIALIZE_PASS_DEPENDENCY(LCSSA)
221 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
222 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
223 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
224 INITIALIZE_PASS_END(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
227 Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognize(); }
229 /// deleteDeadInstruction - Delete this instruction. Before we do, go through
230 /// and zero out all the operands of this instruction. If any of them become
231 /// dead, delete them and the computation tree that feeds them.
233 static void deleteDeadInstruction(Instruction *I, ScalarEvolution &SE,
234 const TargetLibraryInfo *TLI) {
235 SmallVector<Instruction*, 32> NowDeadInsts;
237 NowDeadInsts.push_back(I);
239 // Before we touch this instruction, remove it from SE!
241 Instruction *DeadInst = NowDeadInsts.pop_back_val();
243 // This instruction is dead, zap it, in stages. Start by removing it from
245 SE.forgetValue(DeadInst);
247 for (unsigned op = 0, e = DeadInst->getNumOperands(); op != e; ++op) {
248 Value *Op = DeadInst->getOperand(op);
249 DeadInst->setOperand(op, 0);
251 // If this operand just became dead, add it to the NowDeadInsts list.
252 if (!Op->use_empty()) continue;
254 if (Instruction *OpI = dyn_cast<Instruction>(Op))
255 if (isInstructionTriviallyDead(OpI, TLI))
256 NowDeadInsts.push_back(OpI);
259 DeadInst->eraseFromParent();
261 } while (!NowDeadInsts.empty());
264 /// deleteIfDeadInstruction - If the specified value is a dead instruction,
265 /// delete it and any recursively used instructions.
266 static void deleteIfDeadInstruction(Value *V, ScalarEvolution &SE,
267 const TargetLibraryInfo *TLI) {
268 if (Instruction *I = dyn_cast<Instruction>(V))
269 if (isInstructionTriviallyDead(I, TLI))
270 deleteDeadInstruction(I, SE, TLI);
273 //===----------------------------------------------------------------------===//
275 // Implementation of LIRUtil
277 //===----------------------------------------------------------------------===//
279 // This fucntion will return true iff the given block contains nothing but goto.
280 // A typical usage of this function is to check if the preheader fucntion is
281 // "almost" empty such that generated intrinsic function can be moved across
282 // preheader and to be placed at the end of the preconditiona block without
283 // concerning of breaking data dependence.
284 bool LIRUtil::isAlmostEmpty(BasicBlock *BB) {
285 if (BranchInst *Br = getBranch(BB)) {
286 return Br->isUnconditional() && BB->size() == 1;
291 Value *LIRUtil::getBrCondtion(BasicBlock *BB) {
292 BranchInst *Br = getBranch(BB);
293 return Br ? Br->getCondition() : 0;
296 BasicBlock *LIRUtil::getPrecondBb(BasicBlock *PreHead) {
297 if (BasicBlock *BB = PreHead->getSinglePredecessor()) {
298 BranchInst *Br = getBranch(BB);
299 return Br && Br->isConditional() ? BB : 0;
304 //===----------------------------------------------------------------------===//
306 // Implementation of NclPopcountRecognize
308 //===----------------------------------------------------------------------===//
310 NclPopcountRecognize::NclPopcountRecognize(LoopIdiomRecognize &TheLIR):
311 LIR(TheLIR), CurLoop(TheLIR.getLoop()), PreCondBB(0) {
314 bool NclPopcountRecognize::preliminaryScreen() {
315 const ScalarTargetTransformInfo *STTI = LIR.getScalarTargetTransformInfo();
316 if (STTI->getPopcntHwSupport(32) != ScalarTargetTransformInfo::Fast)
319 // Counting population are usually conducted by few arithmetic instrutions.
320 // Such instructions can be easilly "absorbed" by vacant slots in a
321 // non-compact loop. Therefore, recognizing popcount idiom only makes sense
322 // in a compact loop.
324 // Give up if the loop has multiple blocks or multiple backedges.
325 if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
328 BasicBlock *LoopBody = *(CurLoop->block_begin());
329 if (LoopBody->size() >= 20) {
330 // The loop is too big, bail out.
334 // It should have a preheader containing nothing but a goto instruction.
335 BasicBlock *PreHead = CurLoop->getLoopPreheader();
336 if (!PreHead || !LIRUtil::isAlmostEmpty(PreHead))
339 // It should have a precondition block where the generated popcount instrinsic
340 // function will be inserted.
341 PreCondBB = LIRUtil::getPrecondBb(PreHead);
348 Value *NclPopcountRecognize::matchCondition (BranchInst *Br,
349 BasicBlock *LoopEntry) const {
350 if (!Br || !Br->isConditional())
353 ICmpInst *Cond = dyn_cast<ICmpInst>(Br->getCondition());
357 ConstantInt *CmpZero = dyn_cast<ConstantInt>(Cond->getOperand(1));
358 if (!CmpZero || !CmpZero->isZero())
361 ICmpInst::Predicate Pred = Cond->getPredicate();
362 if ((Pred == ICmpInst::ICMP_NE && Br->getSuccessor(0) == LoopEntry) ||
363 (Pred == ICmpInst::ICMP_EQ && Br->getSuccessor(1) == LoopEntry))
364 return Cond->getOperand(0);
369 bool NclPopcountRecognize::detectIdiom(Instruction *&CntInst,
372 // Following code tries to detect this idiom:
375 // goto loop-exit // the precondition of the loop
378 // x1 = phi (x0, x2);
379 // cnt1 = phi(cnt0, cnt2);
383 // x2 = x1 & (x1 - 1);
390 // step 1: Check to see if the look-back branch match this pattern:
391 // "if (a!=0) goto loop-entry".
392 BasicBlock *LoopEntry;
393 Instruction *DefX2, *CountInst;
394 Value *VarX1, *VarX0;
395 PHINode *PhiX, *CountPhi;
397 DefX2 = CountInst = 0;
400 LoopEntry = *(CurLoop->block_begin());
402 // step 1: Check if the loop-back branch is in desirable form.
404 if (Value *T = matchCondition (LIRUtil::getBranch(LoopEntry), LoopEntry))
405 DefX2 = dyn_cast<Instruction>(T);
410 // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)"
412 if (DefX2->getOpcode() != Instruction::And)
415 BinaryOperator *SubOneOp;
417 if ((SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(0))))
418 VarX1 = DefX2->getOperand(1);
420 VarX1 = DefX2->getOperand(0);
421 SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(1));
426 Instruction *SubInst = cast<Instruction>(SubOneOp);
427 ConstantInt *Dec = dyn_cast<ConstantInt>(SubInst->getOperand(1));
429 !((SubInst->getOpcode() == Instruction::Sub && Dec->isOne()) ||
430 (SubInst->getOpcode() == Instruction::Add && Dec->isAllOnesValue()))) {
435 // step 3: Check the recurrence of variable X
437 PhiX = dyn_cast<PHINode>(VarX1);
439 (PhiX->getOperand(0) != DefX2 && PhiX->getOperand(1) != DefX2)) {
444 // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1
447 for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI(),
448 IterE = LoopEntry->end(); Iter != IterE; Iter++) {
449 Instruction *Inst = Iter;
450 if (Inst->getOpcode() != Instruction::Add)
453 ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1));
454 if (!Inc || !Inc->isOne())
457 PHINode *Phi = dyn_cast<PHINode>(Inst->getOperand(0));
458 if (!Phi && Phi->getParent() != LoopEntry)
461 // Check if the result of the instruction is live of the loop.
462 bool LiveOutLoop = false;
463 for (Value::use_iterator I = Inst->use_begin(), E = Inst->use_end();
465 if ((cast<Instruction>(*I))->getParent() != LoopEntry) {
466 LiveOutLoop = true; break;
481 // step 5: check if the precondition is in this form:
482 // "if (x != 0) goto loop-head ; else goto somewhere-we-don't-care;"
484 BranchInst *PreCondBr = LIRUtil::getBranch(PreCondBB);
485 Value *T = matchCondition (PreCondBr, CurLoop->getLoopPreheader());
486 if (T != PhiX->getOperand(0) && T != PhiX->getOperand(1))
497 void NclPopcountRecognize::transform(Instruction *CntInst,
498 PHINode *CntPhi, Value *Var) {
500 ScalarEvolution *SE = LIR.getScalarEvolution();
501 TargetLibraryInfo *TLI = LIR.getTargetLibraryInfo();
502 BasicBlock *PreHead = CurLoop->getLoopPreheader();
503 BranchInst *PreCondBr = LIRUtil::getBranch(PreCondBB);
504 const DebugLoc DL = CntInst->getDebugLoc();
506 // Assuming before transformation, the loop is following:
507 // if (x) // the precondition
508 // do { cnt++; x &= x - 1; } while(x);
510 // Step 1: Insert the ctpop instruction at the end of the precondition block
511 IRBuilderTy Builder(PreCondBr);
512 Value *PopCnt, *PopCntZext, *NewCount;
514 PopCnt = createPopcntIntrinsic(Builder, Var, DL);
515 NewCount = PopCntZext =
516 Builder.CreateZExtOrTrunc(PopCnt, cast<IntegerType>(CntPhi->getType()));
518 if (NewCount != PopCnt)
519 (cast<Instruction>(NewCount))->setDebugLoc(DL);
521 // If the popoulation counter's initial value is not zero, insert Add Inst.
522 Value *CntInitVal = CntPhi->getIncomingValueForBlock(PreHead);
523 ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal);
524 if (!InitConst || !InitConst->isZero()) {
525 NewCount = Builder.CreateAdd(PopCnt, InitConst);
526 (cast<Instruction>(NewCount))->setDebugLoc(DL);
530 // Step 2: Replace the precondition from "if(x == 0) goto loop-exit" to
531 // "if(NewCount == 0) loop-exit". Withtout this change, the intrinsic
532 // function would be partial dead code, and downstream passes will drag
533 // it back from the precondition block to the preheader.
535 ICmpInst *PreCond = cast<ICmpInst>(PreCondBr->getCondition());
537 Value *Opnd0 = PopCntZext;
538 Value *Opnd1 = ConstantInt::get(PopCntZext->getType(), 0);
539 if (PreCond->getOperand(0) != Var)
540 std::swap(Opnd0, Opnd1);
542 ICmpInst *NewPreCond =
543 cast<ICmpInst>(Builder.CreateICmp(PreCond->getPredicate(), Opnd0, Opnd1));
544 PreCond->replaceAllUsesWith(NewPreCond);
546 deleteDeadInstruction(PreCond, *SE, TLI);
549 // Step 3: Note that the population count is exactly the trip count of the
550 // loop in question, which enble us to to convert the loop from noncountable
551 // loop into a countable one. The benefit is twofold:
553 // - If the loop only counts population, the entire loop become dead after
554 // the transformation. It is lots easier to prove a countable loop dead
555 // than to prove a noncountable one. (In some C dialects, a infite loop
556 // isn't dead even if it computes nothing useful. In general, DCE needs
557 // to prove a noncountable loop finite before safely delete it.)
559 // - If the loop also performs something else, it remains alive.
560 // Since it is transformed to countable form, it can be aggressively
561 // optimized by some optimizations which are in general not applicable
562 // to a noncountable loop.
564 // After this step, this loop (conceptually) would look like following:
565 // newcnt = __builtin_ctpop(x);
568 // do { cnt++; x &= x-1; t--) } while (t > 0);
569 BasicBlock *Body = *(CurLoop->block_begin());
571 BranchInst *LbBr = LIRUtil::getBranch(Body);
572 ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition());
573 Type *Ty = NewCount->getType();
575 PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", Body->begin());
577 Builder.SetInsertPoint(LbCond);
578 Value *Opnd1 = cast<Value>(TcPhi);
579 Value *Opnd2 = cast<Value>(ConstantInt::get(Ty, 1));
581 cast<Instruction>(Builder.CreateSub(Opnd1, Opnd2, "tcdec", false, true));
583 TcPhi->addIncoming(NewCount, PreHead);
584 TcPhi->addIncoming(TcDec, Body);
586 CmpInst::Predicate Pred = (LbBr->getSuccessor(0) == Body) ?
587 CmpInst::ICMP_UGT : CmpInst::ICMP_SLE;
588 LbCond->setPredicate(Pred);
589 LbCond->setOperand(0, TcDec);
590 LbCond->setOperand(1, cast<Value>(ConstantInt::get(Ty, 0)));
593 // Step 4: All the references to the original population counter outside
594 // the loop are replaced with the NewCount -- the value returned from
595 // __builtin_ctpop().
597 SmallVector<Value *, 4> CntUses;
598 for (Value::use_iterator I = CntInst->use_begin(), E = CntInst->use_end();
600 if (cast<Instruction>(*I)->getParent() != Body)
601 CntUses.push_back(*I);
603 for (unsigned Idx = 0; Idx < CntUses.size(); Idx++) {
604 (cast<Instruction>(CntUses[Idx]))->replaceUsesOfWith(CntInst, NewCount);
608 // step 5: Forget the "non-computable" trip-count SCEV associated with the
609 // loop. The loop would otherwise not be deleted even if it becomes empty.
610 SE->forgetLoop(CurLoop);
613 CallInst *NclPopcountRecognize::createPopcntIntrinsic(IRBuilderTy &IRBuilder,
614 Value *Val, DebugLoc DL) {
615 Value *Ops[] = { Val };
616 Type *Tys[] = { Val->getType() };
618 Module *M = (*(CurLoop->block_begin()))->getParent()->getParent();
619 Value *Func = Intrinsic::getDeclaration(M, Intrinsic::ctpop, Tys);
620 CallInst *CI = IRBuilder.CreateCall(Func, Ops);
626 /// recognize - detect population count idiom in a non-countable loop. If
627 /// detected, transform the relevant code to popcount intrinsic function
628 /// call, and return true; otherwise, return false.
629 bool NclPopcountRecognize::recognize() {
631 if (!LIR.getScalarTargetTransformInfo())
634 LIR.getScalarEvolution();
636 if (!preliminaryScreen())
639 Instruction *CntInst;
642 if (!detectIdiom(CntInst, CntPhi, Val))
645 transform(CntInst, CntPhi, Val);
649 //===----------------------------------------------------------------------===//
651 // Implementation of LoopIdiomRecognize
653 //===----------------------------------------------------------------------===//
655 bool LoopIdiomRecognize::runOnCountableLoop() {
656 const SCEV *BECount = SE->getBackedgeTakenCount(CurLoop);
657 if (isa<SCEVCouldNotCompute>(BECount)) return false;
659 // If this loop executes exactly one time, then it should be peeled, not
660 // optimized by this pass.
661 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
662 if (BECst->getValue()->getValue() == 0)
665 // We require target data for now.
666 if (!getDataLayout())
671 LoopInfo &LI = getAnalysis<LoopInfo>();
672 TLI = &getAnalysis<TargetLibraryInfo>();
674 getTargetLibraryInfo();
676 SmallVector<BasicBlock*, 8> ExitBlocks;
677 CurLoop->getUniqueExitBlocks(ExitBlocks);
679 DEBUG(dbgs() << "loop-idiom Scanning: F["
680 << CurLoop->getHeader()->getParent()->getName()
681 << "] Loop %" << CurLoop->getHeader()->getName() << "\n");
683 bool MadeChange = false;
684 // Scan all the blocks in the loop that are not in subloops.
685 for (Loop::block_iterator BI = CurLoop->block_begin(),
686 E = CurLoop->block_end(); BI != E; ++BI) {
687 // Ignore blocks in subloops.
688 if (LI.getLoopFor(*BI) != CurLoop)
691 MadeChange |= runOnLoopBlock(*BI, BECount, ExitBlocks);
696 bool LoopIdiomRecognize::runOnNoncountableLoop() {
697 NclPopcountRecognize Popcount(*this);
698 if (Popcount.recognize())
704 bool LoopIdiomRecognize::runOnLoop(Loop *L, LPPassManager &LPM) {
707 // If the loop could not be converted to canonical form, it must have an
708 // indirectbr in it, just give up.
709 if (!L->getLoopPreheader())
712 // Disable loop idiom recognition if the function's name is a common idiom.
713 StringRef Name = L->getHeader()->getParent()->getName();
714 if (Name == "memset" || Name == "memcpy")
717 SE = &getAnalysis<ScalarEvolution>();
718 if (SE->hasLoopInvariantBackedgeTakenCount(L))
719 return runOnCountableLoop();
720 return runOnNoncountableLoop();
723 /// runOnLoopBlock - Process the specified block, which lives in a counted loop
724 /// with the specified backedge count. This block is known to be in the current
725 /// loop and not in any subloops.
726 bool LoopIdiomRecognize::runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
727 SmallVectorImpl<BasicBlock*> &ExitBlocks) {
728 // We can only promote stores in this block if they are unconditionally
729 // executed in the loop. For a block to be unconditionally executed, it has
730 // to dominate all the exit blocks of the loop. Verify this now.
731 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
732 if (!DT->dominates(BB, ExitBlocks[i]))
735 bool MadeChange = false;
736 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
737 Instruction *Inst = I++;
738 // Look for store instructions, which may be optimized to memset/memcpy.
739 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
741 if (!processLoopStore(SI, BECount)) continue;
744 // If processing the store invalidated our iterator, start over from the
751 // Look for memset instructions, which may be optimized to a larger memset.
752 if (MemSetInst *MSI = dyn_cast<MemSetInst>(Inst)) {
754 if (!processLoopMemSet(MSI, BECount)) continue;
757 // If processing the memset invalidated our iterator, start over from the
769 /// processLoopStore - See if this store can be promoted to a memset or memcpy.
770 bool LoopIdiomRecognize::processLoopStore(StoreInst *SI, const SCEV *BECount) {
771 if (!SI->isSimple()) return false;
773 Value *StoredVal = SI->getValueOperand();
774 Value *StorePtr = SI->getPointerOperand();
776 // Reject stores that are so large that they overflow an unsigned.
777 uint64_t SizeInBits = TD->getTypeSizeInBits(StoredVal->getType());
778 if ((SizeInBits & 7) || (SizeInBits >> 32) != 0)
781 // See if the pointer expression is an AddRec like {base,+,1} on the current
782 // loop, which indicates a strided store. If we have something else, it's a
783 // random store we can't handle.
784 const SCEVAddRecExpr *StoreEv =
785 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
786 if (StoreEv == 0 || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
789 // Check to see if the stride matches the size of the store. If so, then we
790 // know that every byte is touched in the loop.
791 unsigned StoreSize = (unsigned)SizeInBits >> 3;
792 const SCEVConstant *Stride = dyn_cast<SCEVConstant>(StoreEv->getOperand(1));
794 if (Stride == 0 || StoreSize != Stride->getValue()->getValue()) {
795 // TODO: Could also handle negative stride here someday, that will require
796 // the validity check in mayLoopAccessLocation to be updated though.
797 // Enable this to print exact negative strides.
798 if (0 && Stride && StoreSize == -Stride->getValue()->getValue()) {
799 dbgs() << "NEGATIVE STRIDE: " << *SI << "\n";
800 dbgs() << "BB: " << *SI->getParent();
806 // See if we can optimize just this store in isolation.
807 if (processLoopStridedStore(StorePtr, StoreSize, SI->getAlignment(),
808 StoredVal, SI, StoreEv, BECount))
811 // If the stored value is a strided load in the same loop with the same stride
812 // this this may be transformable into a memcpy. This kicks in for stuff like
813 // for (i) A[i] = B[i];
814 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
815 const SCEVAddRecExpr *LoadEv =
816 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getOperand(0)));
817 if (LoadEv && LoadEv->getLoop() == CurLoop && LoadEv->isAffine() &&
818 StoreEv->getOperand(1) == LoadEv->getOperand(1) && LI->isSimple())
819 if (processLoopStoreOfLoopLoad(SI, StoreSize, StoreEv, LoadEv, BECount))
822 //errs() << "UNHANDLED strided store: " << *StoreEv << " - " << *SI << "\n";
827 /// processLoopMemSet - See if this memset can be promoted to a large memset.
828 bool LoopIdiomRecognize::
829 processLoopMemSet(MemSetInst *MSI, const SCEV *BECount) {
830 // We can only handle non-volatile memsets with a constant size.
831 if (MSI->isVolatile() || !isa<ConstantInt>(MSI->getLength())) return false;
833 // If we're not allowed to hack on memset, we fail.
834 if (!TLI->has(LibFunc::memset))
837 Value *Pointer = MSI->getDest();
839 // See if the pointer expression is an AddRec like {base,+,1} on the current
840 // loop, which indicates a strided store. If we have something else, it's a
841 // random store we can't handle.
842 const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Pointer));
843 if (Ev == 0 || Ev->getLoop() != CurLoop || !Ev->isAffine())
846 // Reject memsets that are so large that they overflow an unsigned.
847 uint64_t SizeInBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
848 if ((SizeInBytes >> 32) != 0)
851 // Check to see if the stride matches the size of the memset. If so, then we
852 // know that every byte is touched in the loop.
853 const SCEVConstant *Stride = dyn_cast<SCEVConstant>(Ev->getOperand(1));
855 // TODO: Could also handle negative stride here someday, that will require the
856 // validity check in mayLoopAccessLocation to be updated though.
857 if (Stride == 0 || MSI->getLength() != Stride->getValue())
860 return processLoopStridedStore(Pointer, (unsigned)SizeInBytes,
861 MSI->getAlignment(), MSI->getValue(),
866 /// mayLoopAccessLocation - Return true if the specified loop might access the
867 /// specified pointer location, which is a loop-strided access. The 'Access'
868 /// argument specifies what the verboten forms of access are (read or write).
869 static bool mayLoopAccessLocation(Value *Ptr,AliasAnalysis::ModRefResult Access,
870 Loop *L, const SCEV *BECount,
871 unsigned StoreSize, AliasAnalysis &AA,
872 Instruction *IgnoredStore) {
873 // Get the location that may be stored across the loop. Since the access is
874 // strided positively through memory, we say that the modified location starts
875 // at the pointer and has infinite size.
876 uint64_t AccessSize = AliasAnalysis::UnknownSize;
878 // If the loop iterates a fixed number of times, we can refine the access size
879 // to be exactly the size of the memset, which is (BECount+1)*StoreSize
880 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
881 AccessSize = (BECst->getValue()->getZExtValue()+1)*StoreSize;
883 // TODO: For this to be really effective, we have to dive into the pointer
884 // operand in the store. Store to &A[i] of 100 will always return may alias
885 // with store of &A[100], we need to StoreLoc to be "A" with size of 100,
886 // which will then no-alias a store to &A[100].
887 AliasAnalysis::Location StoreLoc(Ptr, AccessSize);
889 for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
891 for (BasicBlock::iterator I = (*BI)->begin(), E = (*BI)->end(); I != E; ++I)
892 if (&*I != IgnoredStore &&
893 (AA.getModRefInfo(I, StoreLoc) & Access))
899 /// getMemSetPatternValue - If a strided store of the specified value is safe to
900 /// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should
901 /// be passed in. Otherwise, return null.
903 /// Note that we don't ever attempt to use memset_pattern8 or 4, because these
904 /// just replicate their input array and then pass on to memset_pattern16.
905 static Constant *getMemSetPatternValue(Value *V, const DataLayout &TD) {
906 // If the value isn't a constant, we can't promote it to being in a constant
907 // array. We could theoretically do a store to an alloca or something, but
908 // that doesn't seem worthwhile.
909 Constant *C = dyn_cast<Constant>(V);
910 if (C == 0) return 0;
912 // Only handle simple values that are a power of two bytes in size.
913 uint64_t Size = TD.getTypeSizeInBits(V->getType());
914 if (Size == 0 || (Size & 7) || (Size & (Size-1)))
917 // Don't care enough about darwin/ppc to implement this.
918 if (TD.isBigEndian())
921 // Convert to size in bytes.
924 // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see
925 // if the top and bottom are the same (e.g. for vectors and large integers).
926 if (Size > 16) return 0;
928 // If the constant is exactly 16 bytes, just use it.
929 if (Size == 16) return C;
931 // Otherwise, we'll use an array of the constants.
932 unsigned ArraySize = 16/Size;
933 ArrayType *AT = ArrayType::get(V->getType(), ArraySize);
934 return ConstantArray::get(AT, std::vector<Constant*>(ArraySize, C));
938 /// processLoopStridedStore - We see a strided store of some value. If we can
939 /// transform this into a memset or memset_pattern in the loop preheader, do so.
940 bool LoopIdiomRecognize::
941 processLoopStridedStore(Value *DestPtr, unsigned StoreSize,
942 unsigned StoreAlignment, Value *StoredVal,
943 Instruction *TheStore, const SCEVAddRecExpr *Ev,
944 const SCEV *BECount) {
946 // If the stored value is a byte-wise value (like i32 -1), then it may be
947 // turned into a memset of i8 -1, assuming that all the consecutive bytes
948 // are stored. A store of i32 0x01020304 can never be turned into a memset,
949 // but it can be turned into memset_pattern if the target supports it.
950 Value *SplatValue = isBytewiseValue(StoredVal);
951 Constant *PatternValue = 0;
953 // If we're allowed to form a memset, and the stored value would be acceptable
954 // for memset, use it.
955 if (SplatValue && TLI->has(LibFunc::memset) &&
956 // Verify that the stored value is loop invariant. If not, we can't
957 // promote the memset.
958 CurLoop->isLoopInvariant(SplatValue)) {
959 // Keep and use SplatValue.
961 } else if (TLI->has(LibFunc::memset_pattern16) &&
962 (PatternValue = getMemSetPatternValue(StoredVal, *TD))) {
963 // It looks like we can use PatternValue!
966 // Otherwise, this isn't an idiom we can transform. For example, we can't
967 // do anything with a 3-byte store.
971 // The trip count of the loop and the base pointer of the addrec SCEV is
972 // guaranteed to be loop invariant, which means that it should dominate the
973 // header. This allows us to insert code for it in the preheader.
974 BasicBlock *Preheader = CurLoop->getLoopPreheader();
975 IRBuilder<> Builder(Preheader->getTerminator());
976 SCEVExpander Expander(*SE, "loop-idiom");
978 // Okay, we have a strided store "p[i]" of a splattable value. We can turn
979 // this into a memset in the loop preheader now if we want. However, this
980 // would be unsafe to do if there is anything else in the loop that may read
981 // or write to the aliased location. Check for any overlap by generating the
982 // base pointer and checking the region.
983 unsigned AddrSpace = cast<PointerType>(DestPtr->getType())->getAddressSpace();
985 Expander.expandCodeFor(Ev->getStart(), Builder.getInt8PtrTy(AddrSpace),
986 Preheader->getTerminator());
989 if (mayLoopAccessLocation(BasePtr, AliasAnalysis::ModRef,
991 StoreSize, getAnalysis<AliasAnalysis>(), TheStore)){
993 // If we generated new code for the base pointer, clean up.
994 deleteIfDeadInstruction(BasePtr, *SE, TLI);
998 // Okay, everything looks good, insert the memset.
1000 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
1001 // pointer size if it isn't already.
1002 Type *IntPtr = TD->getIntPtrType(DestPtr->getContext());
1003 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtr);
1005 const SCEV *NumBytesS = SE->getAddExpr(BECount, SE->getConstant(IntPtr, 1),
1008 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
1012 Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
1016 NewCall = Builder.CreateMemSet(BasePtr, SplatValue,NumBytes,StoreAlignment);
1018 Module *M = TheStore->getParent()->getParent()->getParent();
1019 Value *MSP = M->getOrInsertFunction("memset_pattern16",
1020 Builder.getVoidTy(),
1021 Builder.getInt8PtrTy(),
1022 Builder.getInt8PtrTy(), IntPtr,
1025 // Otherwise we should form a memset_pattern16. PatternValue is known to be
1026 // an constant array of 16-bytes. Plop the value into a mergable global.
1027 GlobalVariable *GV = new GlobalVariable(*M, PatternValue->getType(), true,
1028 GlobalValue::InternalLinkage,
1029 PatternValue, ".memset_pattern");
1030 GV->setUnnamedAddr(true); // Ok to merge these.
1031 GV->setAlignment(16);
1032 Value *PatternPtr = ConstantExpr::getBitCast(GV, Builder.getInt8PtrTy());
1033 NewCall = Builder.CreateCall3(MSP, BasePtr, PatternPtr, NumBytes);
1036 DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n"
1037 << " from store to: " << *Ev << " at: " << *TheStore << "\n");
1038 NewCall->setDebugLoc(TheStore->getDebugLoc());
1040 // Okay, the memset has been formed. Zap the original store and anything that
1042 deleteDeadInstruction(TheStore, *SE, TLI);
1047 /// processLoopStoreOfLoopLoad - We see a strided store whose value is a
1048 /// same-strided load.
1049 bool LoopIdiomRecognize::
1050 processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
1051 const SCEVAddRecExpr *StoreEv,
1052 const SCEVAddRecExpr *LoadEv,
1053 const SCEV *BECount) {
1054 // If we're not allowed to form memcpy, we fail.
1055 if (!TLI->has(LibFunc::memcpy))
1058 LoadInst *LI = cast<LoadInst>(SI->getValueOperand());
1060 // The trip count of the loop and the base pointer of the addrec SCEV is
1061 // guaranteed to be loop invariant, which means that it should dominate the
1062 // header. This allows us to insert code for it in the preheader.
1063 BasicBlock *Preheader = CurLoop->getLoopPreheader();
1064 IRBuilder<> Builder(Preheader->getTerminator());
1065 SCEVExpander Expander(*SE, "loop-idiom");
1067 // Okay, we have a strided store "p[i]" of a loaded value. We can turn
1068 // this into a memcpy in the loop preheader now if we want. However, this
1069 // would be unsafe to do if there is anything else in the loop that may read
1070 // or write the memory region we're storing to. This includes the load that
1071 // feeds the stores. Check for an alias by generating the base address and
1072 // checking everything.
1073 Value *StoreBasePtr =
1074 Expander.expandCodeFor(StoreEv->getStart(),
1075 Builder.getInt8PtrTy(SI->getPointerAddressSpace()),
1076 Preheader->getTerminator());
1078 if (mayLoopAccessLocation(StoreBasePtr, AliasAnalysis::ModRef,
1079 CurLoop, BECount, StoreSize,
1080 getAnalysis<AliasAnalysis>(), SI)) {
1082 // If we generated new code for the base pointer, clean up.
1083 deleteIfDeadInstruction(StoreBasePtr, *SE, TLI);
1087 // For a memcpy, we have to make sure that the input array is not being
1088 // mutated by the loop.
1089 Value *LoadBasePtr =
1090 Expander.expandCodeFor(LoadEv->getStart(),
1091 Builder.getInt8PtrTy(LI->getPointerAddressSpace()),
1092 Preheader->getTerminator());
1094 if (mayLoopAccessLocation(LoadBasePtr, AliasAnalysis::Mod, CurLoop, BECount,
1095 StoreSize, getAnalysis<AliasAnalysis>(), SI)) {
1097 // If we generated new code for the base pointer, clean up.
1098 deleteIfDeadInstruction(LoadBasePtr, *SE, TLI);
1099 deleteIfDeadInstruction(StoreBasePtr, *SE, TLI);
1103 // Okay, everything is safe, we can transform this!
1106 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
1107 // pointer size if it isn't already.
1108 Type *IntPtr = TD->getIntPtrType(SI->getContext());
1109 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtr);
1111 const SCEV *NumBytesS = SE->getAddExpr(BECount, SE->getConstant(IntPtr, 1),
1114 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
1118 Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
1121 Builder.CreateMemCpy(StoreBasePtr, LoadBasePtr, NumBytes,
1122 std::min(SI->getAlignment(), LI->getAlignment()));
1123 NewCall->setDebugLoc(SI->getDebugLoc());
1125 DEBUG(dbgs() << " Formed memcpy: " << *NewCall << "\n"
1126 << " from load ptr=" << *LoadEv << " at: " << *LI << "\n"
1127 << " from store ptr=" << *StoreEv << " at: " << *SI << "\n");
1130 // Okay, the memset has been formed. Zap the original store and anything that
1132 deleteDeadInstruction(SI, *SE, TLI);